Turbo Product Codes Based on Convolutional Codes
Orhan Gazi and Ali Özgür Yılmaz
In this article, we introduce a new class of product codes
based on convolutional codes, called convolutional
product codes. The structure of product codes enables
parallel decoding, which can significantly increase decoder
speed in practice. The use of convolutional codes in a
product code setting makes it possible to use the vast
knowledge base for convolutional codes as well as their
flexibility in fast parallel decoders. Just as in turbo codes,
interleaving turns out to be critical for the performance of
convolutional product codes. The practical decoding
advantages over serially-concatenated convolutional codes
are emphasized.
Keywords: Iterative decoding, convolutional product
codes, interleaving, puncturing.
Manuscript received Sept. 25, 2005; revised Apr. 26, 2006.
This work was supported in part by the Scientific and Technological Research Council of
Turkey (TUBITAK) under grant 104E027.
Orhan Gazi (phone: + 90 312 284 4500/4015, email: o.gazi@ari.cankaya.edu.tr) is with
Department of Electronic and Communication Engineering, Cankaya University, Ankara,
Turkey.
Ali Özgür Yılmaz (email: aoyilmaz@eee.metu.edu.tr) is with Department of Electrical and
Electronics Engineering, Middle East Technical University, Ankara, Turkey.
ETRI Journal, Volume 28, Number 4, August 2006
I. Introduction
One of the most successful works to approach the Shannon
limit was published in 1993 by C. Berrou, A. Glavieux, and P.
Thitimajshima [1]. They introduced turbo codes, also known as
parallel concatenated convolutional codes (PCCC). Turbo
codes can achieve bit error rate (BER) levels of around 10-5 at
code rates quite close to the corresponding capacity with
reasonable decoding complexity. The use of soft-in soft-out
decoding algorithms was a key to this success. In the last
decade, similar codes such as serially-concatenated
convolutional codes (SCCCs) [2], low-density parity check
codes [3], and block product codes have been extensively
studied. The studies on product codes were initiated by Elias
[4]. Product codes enjoy a high degree of parallelization as
opposed to many other forms of concatenated code structures,
for example, PCCC.
The product codes studied thus far have been constructed
using linear block codes, such as Hamming, extended
Hamming [5], [6], BCH [7], [8], and Reed Solomon [9]
codes. Single parity check (SPC) product codes are studied in
[10]. Three and more dimensional SPC product codes are
studied in [11]. Product codes have also attracted practical
attention lately. Digital signal processing and field-
programmable gate array implementations are studied in [12].
Product codes are traditionally constructed by linear block
codes. Block codes have a trellis structure with a time varying
property [13]. The product code we propose in this paper is
constructed by using time-invariant convolutional codes. Its
component codes’ trellis structure does not vary in time as in
product codes constructed with Hamming, extended
Hamming, BCH, and Reed Solomon block codes. Moreover,
the number of states in the trellis structure of a block code
may grow exponentially with the difference of codeword and
data block lengths [13], whereas the number of states in a
Orhan Gazi et al. 453
convolutional code can be set as desired. The time invariant
trellis structure of convolutional codes makes them more
Encode rows
convenient for implementation. In addition, numerous Data bits Data bits
Row parity
bits
practical techniques such as trellis coded modulation and
puncturing can be simply utilized with convolutional codes as
opposed to linear block codes. A code from the same family
was previously studied for orthogonal frequency-division
multiplexing in [14] but was not analyzed or further Encode columns

elaborated.
Multi-input multi-output (MIMO) techniques are quite
important to enhance the capacity of wireless communication
systems. Space-time trellis codes provide both diversity and Row parity
Data bits
coding gain in MIMO channels and are widely used [15]. bits
Space-time trellis codes usually have time-invariant trellis
structures just like convolutional codes. Thus, a product code Data Row
based on convolutional codes is more suitable for integration Column parity bits column
Parity parity bits
with MIMO channels and poses an alternative to block product Parity
codes.
Due to these advantages of convolutional codes, we propose Fig. 1. Regular product code encoding procedure, where a block
a class of product codes constructed by using convolutional code is used to encode rows and columns.
codes, which we call convolutional product codes (CPCs). In
this paper, we will investigate the factors that affect the
performance of CPCs and leave the issues regarding space Encode rows
time trellis codes to other publications. Data bits
The outline of the paper is as follows. The proposed code
structure and the decoding algorithm for CPCs are given in
section II. The minimum distance of these codes is studied in
section III. In section IV, implementation advantages of CPCs Encode columns
are given. Simulation results are presented in section V.
Concluding remarks are given in section VI.

II. CPC Encoder and Decoder

1. CPC Encoder
A regular product code is constructed by placing the
information bits/symbols into a matrix. The rows and columns
are encoded separately using linear block codes [5]-[8]. This
type of a product encoder is shown in Fig. 1. It is seen from the
figure that the data and parity bits are grouped separately.
In our case, we use convolutional codes instead of linear
block codes to encode rows and columns. This is illustrated in
Fig. 2. When compared to Fig. 1, it is obvious that data and
parity bits are mixed uniformly.
Encoding is performed by using a matrix that determines
how each encoder works. The data to be sent is put into the
matrix. Each row of the matrix is encoded using a
convolutional code. We use the same recursive systematic
convolutional code to encode each row, although different
convolutional codes can be used for this purpose. Once each
Data
Parity
Fig. 2. CPC encoding procedure without an interleaver.
row is encoded, the matrix is sent, if desired, to an interleaver.
Our data matrix dimension is k×k, and the encoded data matrix
dimension is n×n, that is, our code is an (n×n, k×k) code. The
interleaved matrix is coded column-wise. In our simulation we
used the rate 1/2 recursive systematic convolutional code with
the matrix generator (1, 5/7)octal to encode each row and column.
Hence, the overall code rate is 1/4. The general encoding
procedure, which includes any type of interleaver, is illustrated
in Fig. 3.

454 Orhan Gazi et al. ETRI Journal, Volume 28, Number 4, August 2006
 Encode
⎡d1 d 2 ⎤ rows ⎡d1 p1 d 2 p2 ⎤ Interleave ⎡a b c d ⎤ Input data Outer Inner Output data
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ convolutional code Interleaver convolutional code
⎣d 3 d 4 ⎦ ⎣d3 p3 d 4 p4 ⎦ ⎣e f g h ⎦ (row encoder) (column encoder)
Encode
columns
⎡ a b c d ⎤
Fig. 5. SCCC encoding operation.
⎢ ⎥
⎢ pa pb pc pd ⎥
⎢ e f g h ⎥
⎢ ⎥ many log-MAP decoders in-parallel, thus showing smaller
⎢⎣ pe p f p g ph ⎥
⎦ decoding delays. Therefore, we will compare the proposed
Fig. 3. Convolutional product code encoder with any type of CPC structure to that of SCCC.
interleaver (d denotes data bits and p denotes parity bits).

2. CPC Decoder
Convolutional product coded data is multiplexed to a single
stream and binary phase shift key (BPSK) modulated. The
BPSK-modulated signal is passed through an additive white
Gaussian noise channel with double-sided noise power spectral
density N0⁄2, that is, the noise variance is σ 2 = N 0 2 . We
used the log-MAP soft decoding algorithm [16], [17] to
iteratively decode the convolutional product code. Since
columns were encoded last, each column is independently
decoded one by one. The extrinsic information obtained from the
columns is passed to the row decoder after being de-interleaved.
Then, row decoding proceeds; rows are decoded one by one, and
interleaved extrinsic information is passed to the column decoder.
The CPC decoding procedure is depicted in Fig. 4. This
procedure is repeated for a sufficient number of times. The
decoding structure employed in this study is the same as that of
serially-concatenated codes in Fig. 5 [17]. For frames of equal
length, an SCCC decoder uses two log-MAP decoders and
performs quite well at low rates. CPC decoders can utilize
3. Puncturing
Puncturing is a widely used tool to increase the code rate of
convolutional codes [18]. The puncturing operation increases
the rate of a code, but decreases the free distance. This results in
a worse error rate performance compared to a non-punctured
case. We used the puncturing matrix
⎡1 1⎤
⎢1
⎣ 0⎥⎦
to puncture the convolutional component codes.
We studied two cases. In the first case, puncturing is applied
only to the column encoders, resulting in a code rate of 2/3
each. The overall code rate becomes (1/2)×(2/3)=1/3. When
trellis termination is used for rows and columns, a
convolutional code with a slightly smaller overall code rate
(≤ 1 / 3) is produced. In the other case, we apply puncturing to
each row and column encoder, resulting in an increased code
rate of approximately (2/3)×(2/3)=4/9. Simulation results for
punctured convolutional product codes (PCPCs) will be
presented in section V.

Pass it to column
decoders Interleave ⎡ Led1 Lep1 Led 2 Lep 2 ⎤ III. CPC Minimum Distance and its Asymptotic
⎢ ⎥
L L L L
⎣⎢ ed 3 ep3 ed 4 ep 4 ⎦⎥
Performance
The Hamming weight of a binary codeword is defined as
⎡a b c d ⎤ the number of ‘1’s available in the codeword [13]. The
⎢ ⎥ De-interleave Decode each row
⎢ pa pb pc pd ⎥ ⎡d1 p1 d 2 p2 ⎤
⎡a b c d⎤ ⎢ ⎥ using a different minimum distance of a linear code is the minimum Hamming
⎢e f g h ⎥ ⎢ ⎥ ⎣d3 p3 d 4 p4 ⎦
⎢ ⎥ ⎣e f g h⎦ log-MAP decoder
⎢⎣ pe p f p g ph ⎥ weight of all the codewords. The minimum distance plays an
⎦
important role in the code performance. As it gets larger, code
performance improves, especially at high signal-to-noise ratio
Decide on bits
Decode each column
(SNR) values [13]. We assume that dfree is the free distance of
using a different log-MAP the component convolutional codes used in CPCs with trellis
decoder, extract extrinsic
information for the bits a-h Pass it to row decoder
termination. We will investigate the minimum distance of the
CPCs according to the usage of the interleavers.

⎡ Lea Leb Lec Led ⎤ De-interleave

⎢ ⎥ ⎡ Led1 Lep1 Led 2 Lep 2 ⎤ 1. No Interleaving
L L L Leh ⎥ ⎢ ⎥
⎣⎢ ee ef eg ⎦ L L L L
⎣⎢ ed 3 ep3 ed 4 ep 4 ⎦⎥
After the first stage of the CPC encoding operation (row
Fig. 4. Decoding operation of the convolutional product code. encoding), it is obvious that one of the rows of the row-

ETRI Journal, Volume 28, Number 4, August 2006 Orhan Gazi et al. 455
 Encode rows Encode rows
Row data bits X1X1XXX1X1XX1XXX Row data bits X1X1XXX1X1XX1XXX

An S-random interleaver
Encode columns

X X X X X
1 X X X X
X X X X X 1 X X X X
X X X X X X1XXXXXXXXXXXXXX
1 1 1 1 1 1 X X X X
X X X X X 1 X X X X
1 1 1 1 1 X X X X X
X X X X X X X X X X
X X X X X
X 1 X 1 X X X 1 X 1 X X 1 X X X
X X X X X
X X X X X Encode columns
1 1 1 1 1
X X X X X
X X X X X
X X X X X X X X X X
1 1 1 1 1 X X X X X
X X X X X X X X X X
X X X X X
2
X X X X X
Fig. 6. If the column elements are not mixed, d free is preserved. 1 X X X X
1 X X X X
X1XXXXXXXXXXXXXX
1 X X X X
X X X X X
encoded matrix should contain at least dfree number of ‘1’s. This 1 X X X X
1 X X X X
means that there are dfree columns containing at least a single ‘1’ 1 X
1 X
X X
X X
X
X
in the row-encoded matrix. When columns are encoded, there X X
X X
X X
X X
X
X
exists at least dfree number of columns each containing at least 2
2 Fig. 7. d free is not preserved if an S-random interleaver is used
dfree ‘1’s. Hence, in total there are at least d free ‘1’s in the
(‘x’ stands for a single or a group of 0’s).
coded matrix [6]. This is the dmin distance of the CPC whose
component convolutional codes have a trellis termination
constraint. In Figs. 6 and 7, this concept is explained for (1, a single ‘1’ is not necessarily equal to dfree=5. As a result, the CPC
2
5/7)octal component convolutional codes whose free distance is minimum distance is no longer necessarily equal to d free . In fact,
5. In summary, if no interleaver is used, the CPC minimum after an interleaving operation, all the ‘1’s may appear in a single
2
distance is d free . column. This means that the CPC minimum distance is lower
bounded by dfree. We call this type of interleaving full S-random
2. Column S-Random Interleaver interleaving. In Fig. 7, the effect of the full S-random interleaver
2
is illustrated. It is seen from Fig. 7 that when the row-encoded
Both to preserve the d free minimum distance of the CPC,
matrix is S-random interleaved, all the ‘1’s appearing in a row
and to benefit from the interleaving gain, after row encoding
may go to a single column. This verifies that the CPC minimum
we used S-random interleavers for each column, that is, each
distance is lower bounded by dfree.
column is interleaved but different column elements are not
mixed. In this way, we guarantee that dfree number of columns
4. Punctured CPCs
contains a single ‘1’ before column encoding operation. We
call this type of interleaving column S-random interleaving to The puncturing operation decreases the free distance of
distinguish it from regular S-random interleaving. A helical convolutional codes. In our case, we puncture the (1, 5/7)octal
interleaver [19] also does not mix the different column component convolutional code that has dfree=5, that is, an
elements. A helical interleaver and a combination of helical and input sequence ‘0111’ produces minimum Hamming weight
column S-random interleavers will also be considered. codeword ‘00111011’. When the puncturing matrix is applied,
its free distance decreases to dfree=3, that is, deleting every
3. Full S-Random Interleaver second parity bit in a periodic manner, ‘001x101x’ is obtained
from the minimum Hamming weight codeword. Hence, the
If an S-random interleaver is used for all the elements of a CPCs constructed using punctured component convolutional
matrix after row encoding, the number of columns that contain codes have a smaller minimum distance. In fact,

456 Orhan Gazi et al. ETRI Journal, Volume 28, Number 4, August 2006
 2
the minimum distance is equal to d free = 9 if no interleaving
operation is performed or a column S-random interleaver is
used.
5. Asymptotic Performance
If row and column convolutional codes are trellis terminated,
the row and column convolutional codes can be considered as
block codes. Asymptotic performance studies made for block
product codes are also valid for a convolutional product code.
The BER probability of the CPCs can be approximated using
the formula,
N c2, d wc2, d ⎛ 2Es ⎞ to the decoding operation [20]-[21], but also may suffer from
Pb ≈ free free
Q⎜ d 2free ⎟, E s = r 2 Eb ,
k 2 ⎜ N0 ⎟
⎝ ⎠
where dfree is the free distance of the convolutional code used to
construct the convolutional product code, N c , dfree is the number
of convolutional codewords with free distance dfree, k is the length
of the data information frame, and wc , dfree is the average
Hamming weight of the information words that produces
convolutional codewords with Hamming weight dfree. Also, r is
the rate of the component convolutional codes and is equal to 1/2
or 2/3 in our case, and Q is the error function given as
−t 2
1 ∞
∫π e
Q( x) = 2
dt.
2π
The BER approximation in (1) is valid if no interleaver is
used during the CPC encoding operation. If an interleaver is
used, it does not hold anymore.
IV. Practical Implementation Advantages
The implementation advantage of CPC will be discussed
herein with the parameters used in this study. Trellis
termination will be neglected in calculation and will not alter
the results significantly. In SCCC, for a given transmit data
vector of length L, two log-MAP decoders are needed. The first
decoder has a complexity of order O(2L)1) and a time delay of
O(2L). The second decoder has a shorter input, thus it has a
complexity of O(L) and a time delay of O(L). In total, the
complexity is of O(3L) and the time delay is of O(3L). In CPC,
columns are decoded first. The use of separate log-MAP
decoders for each row and column makes parallel processing
operations possible. Each column decoder has a complexity
1) The meaning of O(L) here is different from its conventional usage. By O(L) we mean
that the computation load is proportional to L, i.e., computation amount is approx. kxL where k
is the number of parameters to be computed in a single stage of the code trellis.
ETRI Journal, Volume 28, Number 4, August 2006
of O( L ) and time delay of O( L ). Since these decoders
are run in-parallel, the total column decoding complexity is of
O(2L) but the time delay is of O( L ). Similarly, row
decoding has a total complexity of O(L) and time delay of
O( L ). Hence, although both complexities are the same,
time delays differ very much and bring about an O( L )-
time increase in decoding rate. Hence, the main advantage of
CPCs lies in their suitability for a parallel decoding procedure.
Although there are some proposed methods for the parallel
decoding of SCCCs and PCCCs, these methods usually
propose extra algorithms to solve problems met in parallel
processing. Such algorithms not only bring extra complexity
(1) performance loss. This situation is totally remedied with the
proposed CPCs.
V. Simulation Results
1. Interleaving Effects
A. No Interleaver
In this case, no interleaving operation is performed after row
encoding. Trellis termination bits are added both to rows and
columns. The minimum distance of the CPC is
d min = d free
2
= 25 . Trellis termination bits are necessary to
guarantee d min = d free
2 2
; otherwise dmin is not equal to d free
anymore. The performance graph of this code is shown in Fig.
8. It can be seen from the graph that the performance of this
CPC is not good for low SNR values, although its minimum
distance is large. As is well known, the minimum distance
dominates the performance of the code at high SNR values.
B. Full S-Random Interleaver
After a row-encoding operation, an S-random interleaver (S
= 18) is used. We also simulated a serially-concatenated
convolutional code to compare against CPC due to the
similarity of the code structure and good performance at low
rates. The performance graph is seen in Fig. 8. As seen from
the performance curve, the performance is very good
compared to the cases where interleavers other than S-random
are used. Due to the S-random interleaver used after row
encoding, the minimum distance of the CPC is not necessarily
2
equal to d free . CPC with a full S-random interleaver shows the
best performance at low rates due to the large interleaver gain.
C. Column S-Random Interleaver
To obtain both a better performance than that of the no
interleaver case and to preserve that d min = d free
2
of CPC, we
applied an S-random interleaver (S = 3) to each column
separately. We call such interleaving column S-random
Orhan Gazi et al. 457
 10-1 10-1
C1
CPC No TT
C2
C3 CPC TT
-2
10 10 -2 CPC RT
C4
SCCC
C5
C6
10-3 C7 10-3
BER

10-4 10-4

BER
10-5 10-5

10-6 10-6

10-7 10-7

0.5 1 1.5 2 2.5 3 3.5 0.5 1 1.5 2 2.5

Eb/N0 Eb/N0

C1: No interleaver is used (rate ≈ 1/4)

C2: Theoretical bound (rate ≈ 1/4)
Fig. 9. CPC and SCCC performance graph (frame length = 1024,
C3: Helical interleaver is used (rate ≈ 1/4) iteration number = 12).
C4: Each column is S-random interleaved (column S-random) (rate ≈ 1/4)
C5: Helical + column S-random interleaver is used (rate ≈ 1/4)
C6: Full S-random interleaver is used (rate ≈ 1/4)
C7: SCCC with S-random interleaver (rate ≈ 1/4)
100
Cp1
Cp2
Fig. 8. SCC and CPC performance graph for different interleavers 10-1 Cp3
(iteration number = 12, frame length = 1024). Cp4

10-2
interleaving. Different column elements are not mixed. From
Fig. 8, it can be seen that the performance is better compared to 10-3
BER

the CPC in which no interleaver is used. Its performance is

10-4
worse than the CPC with which a full S-random interleaver is
used after row encoding. The usage of the helical interleaver 10-5

also guarantees that the minimum distance of CPC equals

2 10-6
d free . We also investigated the case in which a helical
interleaver is followed by a column S-random interleaver. It is
seen that such an interleaver results in a slightly better
performance than the one where only a column S-random
interleaver is used during the encoding procedure.
2. Trellis Termination Effects
We simulated three trellis termination cases where trellis
termination bits are added to the rows only (CPC RT), to both
rows and columns (CPC TT), and neither to rows nor to
columns (CPC No TT). Although the addition of trellis
termination bits decreases the code rate, they are critical for
good performance of the convolutional product code as seen in
Fig. 9. The addition of trellis termination bits in a turbo or
serially-concatenated code shows negligible improvement of
the code performance [22]. Without trellis termination, the
performance of the CPC degrades drastically. The performance
graphs are seen in Fig. 9. When only rows are trellis
10-7
0.5 1 1.5 2 2.5 3 3.5 4
Eb/N0
Cp1: SCCC with S-random interleaver (rate ≈ 1/3)
Cp2: CPC with Full S-random interleaver (rate ≈ 1/3)
Cp3: CPC with Full S-random interleaver (rate ≈ 4/9)
Cp4: Each column is S-random interleaved (column S-random) (rate ≈ 4/9)
Fig. 10. Punctured CPC and punctured SCCC performance
graph. CPC rows and columns are trellis terminated.
(frame length = 1024, iteration number = 12).
terminated, a convolutional product code has better
performance at very low Eb/N0 levels. However, the BER
slope decreases at higher Eb/N0 levels when compared to the
case where both rows and columns are trellis terminated. We
see that CPC RT is better than the SCCC and CPC TT at low
Eb/N0 regions. Although it is quite close to BER 10-7, SCCC
seems to have an error curve of higher slope compared to
CPC TT at higher Eb/N0 values.

458 Orhan Gazi et al. ETRI Journal, Volume 28, Number 4, August 2006
3. Puncturing Effects References

Puncturing is first applied only to rows, resulting in a rate 2/3 [1] C. Berrou, A. Glavieux, and P. Thitimajshima, “Near Shannon
CPC. From Fig. 10, it is seen that the performance of the CPC Limit Error-Correcting Coding and Decoding: Turbo-Codes,”
with a full S-random interleaver is good after being punctured. Proc. ICC’93, Geneva, Switzerland, May 1993, pp. 1064-1070.
[2] S. Benedetto, L. Gaggero, R. Garello, and G. Montorsi, “On the
When the puncturing process is applied to both rows and
Design of Binary Serially Concatenated Convolutional Codes,”
columns, it results in a CPC rate of approximately 4/9. From
Proc. VIII Communication Theory Mini-Conf. (CTMC),
Fig. 10, it is seen that the performance becomes very poor for
Vancouver, BC, Canada, June 1999, pp. 32-36.
CPC with a full S-random interleaver. Recall that dmin is not
2
[3] R. G. Gallager, “Low Density Parity Check Codes,” IRE Trans.
necessarily lower bounded by d free when an S-random Inform. Theory, vol. IT-8, Jan. 1962, pp. 21-28.
interleaver is used. Thus, the particular interleaver we used [4] P. Elias, “Error Free Decoding,” IRE Trans. Inform. Theory, vol.
resulted in a low dmin. When d min ≥ d free 2
is ensured by IT-4, Sept. 1954, pp. 29-37.
column S-random interleaving, performance is enhanced [5] E. Hewitt, “Turbo Product Codes for LMDS,” IEEE Radio and
significantly. Wireless Conf., Aug. 1998.
[6] Nam Yul Yu, Young Kim, and Pil Joong Lee, “Iterative Decoding
of Product Codes Composed of Extended Hamming Codes,” 5th
VI. Conclusion
IEEE Symposium on Computers and Communications (ISCC
2000), Antibes, France, 04 – 06 July 2000, pp. 732–737.
In this article, we studied a new class of product codes based
[7] R. M. Pyndiah, “Near-Optimum Decoding of Product Codes:
on convolutional codes. This type of product code has
Block Turbo Codes,” IEEE Trans. Commun., vol. 46, no. 8, Aug.
component codes with a time invariant trellis structure, as
1998, pp. 1003-1010
opposed to product codes constructed with linear block codes
[8] T. Shohon, Y. Soutome, and H. Ogiwara, “Simple Computation
(Hamming, BCH, Reed Solomon, and so on). Hence, CPC
Method of Soft Value for Iterative Decoding of Product Code
may be more favorable for implementation than linear block
Composed of Linear Block Code,” IEIC Trans. Fundamentals,
product codes. When compared to serially-concatenated
vol. E82-A, no. 10, Oct. 1999, pp. 2199–2203.
convolutional codes, it exhibits comparable BER levels which
[9] Omar Aitsab and Ramesh Pyndiah, “Performance of Reed
are of practical interest.
Solomon Block Turbo Codes,” Proc. IEEE LOBECOM’ 96
We investigated the effects of different interleavers on the
Conf., London, U.K., vol. 1/3, Nov. 1996, pp. 121-125.
performance of CPCs. It was seen that CPCs are outperformed
[10] David Rankin and T. Aaron Gulliver, “Single Parity Check
by other codes unless good interleavers are used. We proposed
Product Codes,” IEEE Trans. Commun., vol. 49, no. 8, Aug.
interleaving methods to preserve the greatest minimum
2001, pp. 1354-1362.
distance of CPCs. It is seen that the performance of a CPC is
[11] D. M. Rankin and T. A. Gulliver, “Randomly Interleaved Single
best at low rates when a full S-random interleaver is used.
Parity Check Product Codes,” Proc. IEEE Int Symp. on Inform.
Column S-random interleavers are much better for punctured
Theory, June 2000, pp. 88.
CPCs.
[12] A. Goalic and R. Pyndiah, “Real Time Turbo Decoding of
Currently, we are investigating the effects of various
Product Codes on a Digital Signal Processor,” Int. Symposium on
interleavers and the incorporation of trellis coded modulation in
Turbo Codes and Related Topics, Brest, Sept. 1997, pp. 624-628.
row and column encoding. Since CPCs employ matrices in
[13] S. Lin, D. Costello, Error Control Coding, Prentice Hall, 2004.
encoding, it can be easily extended to multi-carrier modulation
[14] F. Sanzi and S. ten Brink, “Iterative Channel Estimation and
where the vertical dimension can correspond to the sub-carriers.
Decoding with Product Codes in Multicarrier Systems,” IEEE
The approach presented here can be successfully extended to
VTS Fall VTC2000 52nd Vehicular Technology Conf., Boston,
space-time trellis coding. Thus, our future studies will also
MA, USA, Sep. 2000, pp. 1388–1344.
include a joint structure for CPCs and MIMO space-time
[15] V. Tarokh V, N. Seshadri, and A. R. Calderbank, “Space-Time
frequency codes.
Codes for High Data Rate Wireless Communication:
Performance Criterion and Code Construction,” IEEE Trans.
Acknowledgement Inform. Theory, vol. 44, no. 2, Mar. 1998, pp. 744-765.
[16] J. Hagenauer, E. Offer, and L. Papke, “Iterative Decoding of
The authors would like to thank the anonymous reviewers Binary Block and Convolutional Codes,” IEEE Trans. Inform.
for their valuable comments and helpful corrections. Theory, vol. 42, no. 2, Mar. 1996, pp. 429–445.

ETRI Journal, Volume 28, Number 4, August 2006 Orhan Gazi et al. 459
[17] S. Benedetto, D. Divsalar, G. Montorsi, and F. Pollara, “Serially
Concatenation of Interleaved Codes: Design and Performance
Analysis,” IEEE Trans. Inform. Theory, vol. 44, May 1998, pp.
909-926.
[18] J. Hagenauer, “Rate-Compatible Punctured Convolutional Codes
(RCPC Codes) and their Applications,” IEEE Trans. Inform.
Theory, vol. 44, no. 3, May 1998, pp. 909- 926.
[19] Branka Vucetic and Jinhong Yuan, Turbo Codes: Principles and
Applications, Kluwer Academic Publishers, May 2000.
[20] A. Tarable, S. Benedetto, and G. Montorsi, “Mapping Interleaving
Laws to Parallel Turbo and LDPC Decoder Architectures,” IEEE
Transactions on Information Theory, vol. 50, no. 9, Sep. 2004, pp.
2002–2009.
[21] Seokhyun Yoon and Yeheskel Bar-Ness, “A Parallel MAP
Algorithm for Low Latency Turbo Decoding,” IEEE
Communication Letters, vol. 6, no. 7, July 2002, pp. 288–290.
[22] P. Robertson, “Illuminating the Structure of Parallel Concatenated
Recursive (TURBO) Codes,” Proc. GLOBECOM’94, San
Francisco, CA, Nov. 1994, pp. 1298-1303.
460 Orhan Gazi et al.
Orhan Gazi received the BS and MS
degrees in electrical and electronics engineering
from Middle East Technical University, Ankara,
Turkey in 1996 and 2001. Since 2001, he has
been working towards the PhD degree. His
research includes error control coding and signal
processing. He is currently employed as an
Instructor in Cankaya University, where he delivers lectures at the
undergraduate level in electrical engineering.
Ali Özgür Yılmaz received the BS degree in
electrical engineering from the University of
Michigan, Ann Arbor in 1999. He received the
MS and PhD degrees at the same school in
2001 and 2003. He has been an Assistant
Professor with Middle East Technical
University, Turkey since September 2003. His
research interests include high rate mobile communication systems,
cooperative communications, and radar signal processing.
ETRI Journal, Volume 28, Number 4, August 2006